1

A comparison of classification techniques for seismic facies recognition

Tao Zhao1, Vikram Jayaram2, Atish Roy3, and Kurt J. Marfurt1

1The University of Oklahoma, ConocoPhillips School of Geology and Geophysics

2Oklahoma Geological Survey

3BP America

tao-zhao@ou.edu, vjayaram@ou.com, atish.roy@bp.com, and kmarfurt@ou.edu

Corresponding author:

Tao Zhao

The University of Oklahoma, ConocoPhillips School of Geology and Geophysics

710 Sarkeys Energy Center

100 East Boyd Street

Norman, OK 73019

tao-zhao@ou.edu

2

ABSTRACT

During the past decade, the size of 3D seismic data volumes and the number of seismic

attributes have increased to the extent that it difficult, if not impossible, for interpreters to examine

every seismic line and time slice. To address this problem, several seismic facies classification

algorithms including k-means, self-organizing maps, generative topographic mapping, support

vector machines, Gaussian mixture models, and artificial neural networks have been successfully

used to extract features of geologic interest from multiple volumes. Although well documented in

the literatures, the terminology and complexity of these algorithms may bewilder the average

seismic interpreter, and few papers have applied these competing methods to same data volume.

In this paper, we review six commonly used algorithms and apply them to a single 3D seismic data

volume acquired over the Canterbury Basin, offshore New Zealand, where one of the main

objectives is to differentiate the architectural elements of a turbidite system. Not surprisingly, the

most important parameter in this analysis is the choice of the correct input attributes, which in turn

depends on careful pattern recognition by the interpreter. We find that supervised learning methods

provide accurate estimates of the desired seismic facies, while unsupervised learning methods also

highlight features that may otherwise be overlooked.

Keywords: pattern recognition, machine learning, classification, seismic interpretation, reservoir

characterization

3

INTRODUCTION

In 2015, pattern recognition has become part of everyday life. Amazon or Alibaba analyzes

the clothes you buy, Google analyzes your driving routine, and your local grocery store knows the

kind of cereal you eat in the morning. “Big data” and “deep learning algorithms” are being

analyzed by big companies and big government, attempting to identify patterns in our spending

habits and the people with whom we associate.

Successful seismic interpreters are experts at pattern recognition, identifying features such

as channels, mass transport complexes, and collapse features where our engineering colleagues

only see wiggles. Our challenge as interpreters is that the data volumes we need to analyze keep

growing in size and dimensionality, while the number of experienced interpreters has remained

relatively constant. One solution to this dilemma is for these experienced interpreters to teach their

skills to the next generation of geologists and geophysicists, either through traditional or on the

job training. An alternative and complimentary solution is for these experienced interpreters to

teach theirs skills to a machine. Turing (1950), whose scientific contributions and life has recently

been popularized in a movie, asked whether “Machines can think?” Whether machines will ever

be able to think is a question for scientists and philosophers to answer (e.g. Eagleman, 2012), but

machines can be taught to perform repetitive tasks, and even to unravel the relationships that

underlay repetitive patterns, in an area called “Machine Learning”.

25 years ago, skilled interpreters delineated seismic facies on a suite of 2D lines by visually

examining seismic waveforms, frequency, amplitude, phase, and geometric configurations. Facies

would then be posted on a map and hand contoured to generate a seismic facies map. With the

introduction of 3D seismic data and volumetric attributes, such analysis has become both more

4

quantitative and more automated. In this tutorial, we focus on classification (also called clustering)

on large 3D seismic data whereby like patterns in the seismic response (seismic facies) are assigned

similar values. Much of the same technology can be used to define specific rock properties, such

as brittleness, TOC, or porosity. Pattern recognition and clustering is common to many industries,

from using cameras to identify knotholes in plywood production to tracking cell phone

communications to identify potential narcotics traffickers. The workflow is summarized in the

classic textbook by Duda et al. (2001) displayed in Figure 1. In this figure, “sensing” consists of

seismic, well log, completion, and production measurements. For interpreters “segmentation” will

usually mean focusing on a given stratigraphic formation or suite of formations. Seismic data lose

both temporal and lateral resolution with depth, such that a given seismic facies changes its

appearance, or is nonstationary, as we go deeper in the section. The number of potential facies also

increases as we analyze larger vertical windows incorporating different depositional environments,

making classification more difficult. For computer assisted facies classification, “feature

extraction” means attributes, be they simple measurements of amplitude and frequency, geometric

attributes that measure reflector configurations, or more quantitative measurements of lithology,

fractures, or geomechanical properties provided by prestack inversion and azimuthal anisotropy

analysis. “Classification” assigns each voxel to one of a finite number of classes (also called

clusters), each of which represents a seismic facies that may or may not correspond to a geological

facies. Finally, using validation data, the interpreter makes a “decision” that determines whether a

given cluster represents a unique seismic facies, if it should be lumped in other clusters having a

somewhat similar attribute expression, or whether it should be further subdivided, perhaps through

the introduction of additional attributes.

5

Pattern recognition and classification of seismic features is fundamental to human based

interpretation, where our job may be as “simple” as identifying and picking horizons and faults, or

more advanced such as the delineation of channels, mass transport complexes, carbonate buildups,

or potential gas accumulations. The use of computer-assisted classification began soon after the

development of seismic attributes in the 1970s (Balch, 1971; Taner et al., 1979), with the work by

Sonneland (1983) and Justice et al. (1985) being two of the first. K-means (Forgy, 1965; Jancey,

1966) was one of the earliest clustering algorithms developed, and was quickly applied by service

companies and today is common to almost all interpretation software packages. K-means is an

unsupervised learning algorithm in that the interpreter provides no prior information other than the

selection of attributes and the number of desired clusters.

Barnes and Laughlin (2002) reviewed several unsupervised clustering techniques,

including k-means, fuzzy clustering, and self-organizing maps (SOM). Their primary finding was

that the clustering algorithm used was less important than the choice of attributes used. Among the

clustering algorithms, they favored SOM since there is topologically ordered mapping of the

clusters with similar clusters lying adjacent to each other on a manifold and in the associated latent

space. In our examples, a “manifold” is a deformed 2D surface that best fits the distribution of n

attributes lying in an n-dimensional data space. The clusters are then mapped to a simpler 2D

rectangular “latent” (Latin for “hidden”) space upon which the interpreter can either interactively

define clusters or simply map the projections onto a 2D colorbar. A properly chosen latent space

can help identify data properties that are otherwise difficult to observe in the original input space.

Coleou et al.’s (2003) seismic “waveform classification” algorithm is implemented using SOM,

where the “attributes” are seismic amplitudes that lie on a suite of 16 phantom horizon slices. Each

(x,y) location in the analysis window results provides a 16-dimensional vector of amplitudes. When

6

plotted one element after the other, the mean of each cluster in 16-dimensional space looks like a

waveform. These waveforms lie along a 1D deformed string (the manifold) that lies in 16D. This

1D string is then mapped to a 1D line (the latent space) which in turn is mapped against a 1D

continuous color bar. The proximity of like waveforms to each other on the manifold and latent

spaces results in similar seismic facies appearing as similar colors. Coleou et al. (2003) generalized

their algorithm to attributes other than seismic amplitude, constructing vectors of dip magnitude,

coherence, and reflector parallelism. Strecker and Uden (2002) were perhaps the first to use 2D

manifolds and 2D latent spaces with geophysical data, using multidimensional attribute volumes

to form N-dimensional vectors at each seismic sample point. Typical attributes included envelope,

bandwidth, impedance, AVO slope and intercept, dip magnitude, and coherence. These attributes

were projected onto a 2D latent space and their results plotted against a 2D color table. Gao (2007)

applied a 1D SOM to GLCM texture attributes to map seismic facies offshore Angola.

Overdefining the clusters with 256 prototype vectors, he then used 3D visualization and his

knowledge of the depositional environment to map the “natural” clusters. These natural clusters

were then calibrated using well control, giving rise to what is called a posteriori supervision. Roy

et al. (2013) built on these concepts and developed an SOM classification workflow of multiple

seismic attributes computed over a deep-water depositional system. They calibrated the clusters a

posteriori using classical principles of seismic stratigraphy on a subset of vertical slices through

the seismic amplitude. A simple but very important innovation was to project the clusters onto a

2D nonlinear Sammon space (Sammon, 1969). This projection was then colored using a

gradational 2D color-scale like that of Matos et al. (2009) thus facilitating the interpretation. Roy

et al. (2013) introduced a Euclidean distance measure to correlate predefined unsupervised clusters

to average data vectors about interpreter defined well log facies.

7

Generative topographic mapping (GTM) is a more recent unsupervised classification

innovation, providing a probabilistic representation of the data-vectors in the latent space (Bishop

et al., 1998). There has been very little work on the application of GTM technique to seismic data

and exploration problems. Wallet et al. (2009) are probably the first to apply the GTM technique

to seismic data, using a suite of phantom horizon slices through a seismic amplitude volume

generating a “waveform classification”. While generating excellent images, Roy et al. (2013,

2014) found the introduction of well control to SOM classification to be somewhat limited, and

instead applied generative topographic mapping (GTM) to Mississippian tripolitic chert reservoir

in the Midcontinent USA and a carbonate wash play in the Sierra Madre Oriental of Mexico. They

found that GTM provided not only the most likely cluster associated with a given voxel, but also

the probability that that voxel belongs each of clusters, providing a measure of confidence or risk

in the prediction.

K-means, SOM, and GTM are all unsupervised learning techniques, where the clustering

is driven only by the choice of input attributes and the number of desired clusters. If we wish to

teach the computer to mimic the facies identification previously chosen by a skilled interpreter, or

link seismic facies to electro-facies interpreted using wireline logs, we need to introduce

“supervision” or external control to the clustering algorithm. The most popular means of

supervised learning classification are based on artificial neural networks (ANN). Meldahl et al.

(1999) used seismic energy and coherence attributes coupled with interpreter control (picked seed

points) to train a neural network to identify hydrocarbon chimneys. West et al. (2002) used a

similar workflow where the objective was seismic facies analysis of a channel system and the input

attributes were textures. Corradi et al. (2009) used GLCM (gray level co-occurrence matrix)

8

textures and ANN, with controls based on wells and skilled interpretation of some key 2D vertical

slices to map sand, evaporate, and sealing vs. non-sealing shale facies offshore west Africa.

Support vector machine (SVM, where the word “machine” is due to Turing’s (1950)

mechanical decryption machine) is a more recent introduction to (e.g. Li and Castagna, 2004;

Kuzma and Rector, 2004, 2005; Zhao et al., 2005; Al-Anazi and Gates, 2010). Originating from

maximum margin classifiers, SVMs have gained great popularity for solving pattern classification

and regression problems since the concept of a “soft margin” was first introduced by Cortes and

Vapnik (1995). SVMs map the N-dimensional input data into a higher dimensional latent (often

called feature) space, where clusters can be linearly separated by hyperplanes. Detailed description

on SVMs can be found in Cortes and Vapnik (1995), Cristianini and Shawe-Taylor (2000), and

Schölkopf and Smola (2002). Li and Castagna (2004) used SVM to discriminate alternative AVO

responses while Zhao et al. (2014) and Zhang et al. (2015) used a variation of SVM using

mineralogy logs and seismic attributes to predict lithology and brittleness in a shale resource play.

We begin our paper by providing a summary of the more common clustering techniques

used in seismic facies classification, emphasizing their similarities and differences. We start from

the unsupervised learning k-means algorithm, progress through projections onto principal

component hyperplanes, and end with projections onto SOM and GTM manifolds, which are

topological spaces that resemble Euclidean space near each point. Next, we provide a summary of

supervised learning techniques including artificial neural networks and support vector machines.

Given these definitions, we apply each of these methods to identify seismic facies in the same data

volume acquired in the Canterbury Basin of New Zealand. We conclude with a discussion on the

advantages and limitations of each method and areas for future algorithm development and

9

workflow refinement. At the very end, we provide an appendix containing some of the

mathematical details to better quantify how each algorithm works.

REVIEW OF UNSUPERVISED LEARNING CLASSIFICATION TECHNIQUES

Crossplotting

Crossplotting one or more attributes against each other is an interactive and perhaps the

most common clustering technique. In its simplest implementation, one computes and then

displays a 2D histogram of two attributes. In most software packages, the interpreter then identifies

a cluster of interest and draws a polygon around it. While several software packages allow

crossplotting of up to three attributes, crossplotting more than three attributes quickly becomes

intractable. One workflow to address this visualization limitation is to first project a high number

of attributes onto the first two or three eigenvectors, and then crossplot the principal components.

Principal components will be discussed later in the section on projection methods.

K-means classification

K-means (MacQueen, 1967) is perhaps the simplest clustering algorithm and is widely

available in commercial interpretation software packages. The method is summarized in the

cartoons shown in Figure 2. One drawback of the method is that the interpreter needs to define

how many clusters reside in the data. Once the number of clusters is defined, the cluster means or

centers are defined either on a grid or randomly to begin the iteration loop. Since attributes have

different units of measurement (e.g. Hz for peak frequency, 1/km for curvature, and mV for RMS

amplitude) the distance of each data point to the current means are computed by scaling the data

by the inverse of the covariance matrix, giving us the “Mahalanobis” distance (see Appendix).

10

Each data point is then assigned to the cluster to whose mean it is closest. Once assigned, new

cluster means are computed from the newly assigned data clusters and the process repeated. If

there are Q clusters, the process will converge in about Q iterations.

K-means is fast and easy to implement. Unfortunately, the clustering has no structure such

that there is no relationship between the cluster numbering (and therefore coloring) and the

proximity of one cluster to another. This lack of organization can result in similar facies appearing

in totally different colors, confusing the interpretation. Tuning the number of clusters to force

similar facies into the same cluster is a somewhat tedious procedure that also decreases the

resolution of the facies map.

Projection Techniques

Although not defined this way in the pattern recognition literature, since this is a tutorial,

we will lump the following methods, principal component analysis (PCA), self-organizing maps,

and generative topographic maps together and call them “projection techniques”. Projection

techniques project data residing in a higher dimensional space (say a 5D space defined by five

attributes) onto a lower dimensional space (say a 2D plane or deformed 2D surface). Once

projected, the data can be clustered in that space by the algorithm (such as SOM) or interactively

clustered by the interpreter by drawing polygons (routine for PCA, and our preferred analysis

technique for both SOM and GTM).

Principal Component Analysis

Principal component analysis is widely used to reduce the redundancy and excess

dimensionality of the input attribute data. Such reduction is based on the assumption that most of

11

the signals are preserved in the first few principle components (eigenvectors), while the last

principal components contain uncorrelated noise. In this tutorial, we will use PCA as the first

iteration of the SOM and GTM algorithms. Many workers use PCA to reduce redundant attributes

into “meta attributes” to simplify the computation. The first eigenvector is a vector in N-

dimensional attribute space that best represents the attribute patterns in the data. Cross-correlating

(projecting) the N-dimensional data against the first eigenvector at each voxel gives us the first

principal component volume. If we scale the first eigenvector by the first principal component and

subtract it from the original data vector, we obtain a residual data vector. The second eigenvector

is that vector that best represents the attribute patterns in this residual. Cross-correlating

(projecting) the second eigenvector against either the original data or residual data vector at each

voxel gives us the second principal component volume. This process continues for all N-

dimensions resulting in N eigenvectors and N principal components. In this paper, we will limit

ourselves to the first two eigenvectors which thus define the plane that least-squares fits the N-

dimensional attribute data. Figure 3c shows a numerical example of the first two principle

components defining a plane in a 3-dimensional data space.

Self-organizing maps

While many workers (e.g. Coleou et al., 2003) describe SOM as a type of neural network,

for the purposes of this tutorial, we prefer to describe SOM as a manifold projection technique.

Kohonen (1982) SOM, originally developed for gene pattern recognition, is one of the most

popular classification techniques, and it has been implemented in at least four commercial software

packages for seismic facies classification. The major advantage of SOM over k-means is that the

clusters residing on the deformed manifold in N-dimensional data space are directly mapped to a

12

rectilinear or otherwise regularly gridded latent space. We provide a brief summary of the

mathematical formulations of the SOM implementation used in this study in the Appendix.

Although SOM is one of the most popular classification technique there are several

limitations to the SOM algorithm. First, the choice of neighborhood function at each iteration is

subjective, with different choices resulting in different solutions. Second, the absence of a

quantitative error measure does not let us know whether the solution has converged to an

acceptable level, thus providing confidence in the resulting analysis. Third, while we find the most

likely cluster for a given data vector, we have no quantitative measure of confidence in the facies

classification, and no indication if the vector could be nearly as well represented by other facies.

Generative topographic mapping

GTM is a nonlinear dimensionality reduction technique that provides a probabilistic

representation of the data vectors on a lower L-dimensional deformed manifold that is in turn

mapped to an L-dimensional latent space. While SOM seeks the node or prototype vector that is

closest to the randomly chosen vector from the training or input dataset, in GTM each of the nodes

lying on the lower dimensional manifold provides some mathematical support to the data and is

considered to be to some degree “responsible” for the data vector (Figure 4). The level of support

or “responsibility” is modeled with a constrained mixture of Gaussians. The model parameter

estimations are determined by maximum likelihood using the Expectation Maximization (EM)

algorithm (Bishop et al., 1998).

Because GTM theory is deeply rooted in probability, it can also be used in modern risk

analysis. We can extend the GTM application in seismic exploration by projecting the mean

posterior probabilities of a particular window of multiattribute data (say, about a producing well)

13

onto the 2D latent space. By project the data vector at any given voxel onto the latent space, we

obtain a probability estimates of whether it falls into the same category (Roy et al., 2014). We thus

have a probabilistic estimate of how similar any data vector is to attribute behavior (and hence

facies) about a producing or non-producing well of interest.

Other Unsupervised Learning Methods

There are many other unsupervised learning techniques, several of which were evaluated

by Barnes and Laughlin (2002). We do not currently have access to software to apply independent

component analysis and Gaussian mixture models to our seismic facies classification problem, but

mention them as possible candidates.

Independent component analysis

Like PCA, independent component analysis (ICA) is a statistical technique used to project

a set of N-dimensional vectors onto a smaller L-dimensional space. Unlike PCA which is based on

Gaussian statistics, whereby the first eigenvector best represents the variance in the

multidimensional data, ICA attempts to project data onto subspaces that result in non-Gaussian

distributions which are then easier to separate and visualize. Honorio et al. (2014) successfully

apply ICA to multiple spectral components to delineate architectural elements of an offshore Brazil

carbonate terrain. Both PCA and ICA are commonly used to reduce a redundant set of attributes

to form a smaller set of independent meta-attributes (e.g. Gao, 2007).

Gaussian mixture models

Gaussian mixture model (GMM), are parametric models of probability distributions which

can provide greater flexibility and precision in modeling than traditional unsupervised clustering

14

algorithms. Lubo et al. (2014) apply this technique to a suite of well logs acquired over Horseshoe

Atoll, west Texas, to generate different lithologies. These GMM lithologies are then used to

calibrate 3D seismic prestack inversion results to generate a 3D rock property model. At present,

we do not know of any GMM algorithms applied to seismic facies classification using seismic

attributes as input data.

REVIEW OF SUPERVISED LEARNING CLASSIFICATION TECHNIQUES

Artificial Neural Networks

Artificial neural networks can be used in both unsupervised and supervised mulitattribute

analysis (van der Baan and Jutten, 2000). The multilayer perceptron (MLP) and the radial basis

function (RBF) are two popular types of neural networks used in supervised learning. Probabilistic

neural network, PNN, which also uses radial basis functions, forms the basis of additional neural

network geophysical applications. In terms of network architecture, the supervised algorithms are

feed-forward networks. In contrast, the unsupervised SOM algorithm described earlier is a

recurrent (or feed-backward) network. An advantage of feed-forward networks over SOMs is the

ability to predict both continuous values (such as porosity) as well as discrete values (such as facies

class number). Applications of neural networks can be found in seismic inversion (Rӧth and

Tarantola, 1994), well log prediction from other logs (Huang et al., 1996; Lim, 2005), waveform

recognition (Murat and Rudman, 1992), seismic facies analysis (West et al., 2002), and reservoir

property prediction using seismic attributes (Yu et al., 2008; Zhao and Ramachandran, 2013). For

the last application listed above, however, due to the resolution difference between seismic and

well logs, structural and lithologic variation of inter-well points, and the highly nonlinear relation

between these two domains, achieving a convincing prediction result can be challenging. In this

15

case, geostatistical methods such as Bayesian analysis can be used jointly to provide a probability

index, giving interpreters an estimate of how much confidence they should have in the prediction.

Artificial neural networks are routinely used in the exploration and production industry.

ANN provides a means to correlate well measurements such as gamma ray logs to seismic

attributes (e.g. Verma, 2012) where the underlying relationship is a function of rock properties,

depositional environment, and diagenetic alteration. Although it has produced reliable

classification in many applications during its service, defects such as converging to local minima

and difficult in parameterization are not negligible. In both industrial and scientific applications,

we prefer a constant and robust classifier once the training vectors and model parameters have

been determined. This leads to the more recent supervised learning technique developed in the late

20th century, the support vector machines.

Support Vector Machines

The basic idea of SVMs is straightforward. First, we transform the training data vectors

into a still higher dimensional “feature” space using nonlinear mapping. Then we find a hyperplane

in this feature space that separates the data into two classes with an optimal “margin”. The concept

of a margin is defined to be the smallest distance between the separation hyperplane (commonly

called a decision boundary) and the training vectors (Bishop, 2006) (Figure 5). An optimal margin

balances two criteria: maximizing the margin, thereby giving the classifier the best generalization,

and minimizing the number of misclassified training vectors if the training data are not linearly

separable. The margin can also be described as the distance between the decision boundary and

two hyperplanes defined by the data vectors which have the smallest distance to the decision

boundary. These two hyperplanes are called the “plus-plane” and the “minus-plane”. The vectors

16

which lie exactly on these two hyperplanes mathematically define or “support” them and are called

support vectors. Tong and Koller (2002) show that the decision boundary is dependent solely on

the support vectors, resulting in the name “support vector machines”.

SVMs can be used in either a supervised or in a semi-supervised learning mode. In contrast

to supervised learning, semi-supervised learning defines a learning process that utilizes both

labeled and unlabeled vectors. When there are a limited number of interpreter classified data

vectors, the classifier may not act well due to insufficient training. In semi-supervised training,

some of the nearby unclassified data vectors are automatically selected and classified based on a

distance measurement during the training step, as in an unsupervised learning process. These

vectors are then used as additional training vectors (Figure 6), resulting in a classifier that will

perform better for the specific problem. The generalization power is sacrificed by using unlabeled

data. In this tutorial we focus on SVM; however, the future of semi-supervised SVM in

geophysical applications is quite promising.

Proximal Support Vector Machines

Proximal support vector machine (PSVM) (Fung and Mangasarian, 2001, 2005) is a recent

variant of SVM, which, instead of looking for a separating plane directly, builds two parallel planes

that approximate two data classes; the decision-boundary then falls between these two planes

(Figure 7). Other researchers have found that PSVM provides comparable classification

correctness to standard SVM but at considerable computational savings (Fung and Mangasarian,

2001, 2005; Mangasarian and Wild, 2006). In this tutorial, we use PSVM as our implementation

of SVM. Details on the PSVM algorithm are provided in the Appendix.

17

We may face problems in seismic interpretation that are linearly inseparable in the original

input multidimensional attribute space. In SVM, we map the data vectors into a higher dimensional

space where they become linearly separable (Figure 8), where the increase in dimensionality may

result in significantly increased computational cost. Instead of using an explicit mapping function

to map input data into a higher dimensional space, PSVM achieves the same goal by manipulating

a kernel function in the input attribute space. In our implementation, we use a Gaussian kernel

function, but in principal many other functions can be used (Shawe-Taylor and Cristianini, 2004).

SVM can be used either as a classifier or as a regression operator. Used as a regression

operator, SVM is capable of predicting petrophysical properties such as porosity (Wong et al.,

2005), 𝑉𝑝, 𝑉𝑠 and density (Kuzma and Rector, 2004), and permeability (Al-Anazi and Gates, 2010;

Nazari et al., 2011). In all such applications, SVM shows comparable or superior performance to

neural networks with respect to prediction error and training cost. When used as a classifier, SVM

is suitable in predicting lithofacies (Al-Anazi and Gates, 2010; Torres and Reveron, 2013; Wang

et al., 2014; Zhao et al., 2014) or pseudo rock properties (Zhang et al., 2015), either from well log

data, core data, or seismic attributes.

GEOLOGIC SETTING

In this tutorial we utilized the Waka-3D seismic survey acquired over the Canterbury

Basin, offshore New Zealand, generously made public by New Zealand Petroleum and Minerals.

Readers can request this data set through their website for research purposes. Figure 9 shows the

location of this survey, where the red rectangle corresponds to time slices shown in subsequent

figures. The study area lies on the transition zone of continental slope and rise, with abundance of

paleocanyons and turbidite deposits of Cretaceous and Tertiary ages. These sediments are

18

deposited in a single, tectonically driven transgressive – regressive cycle (Uruski, 2010). Being a

very recent and underexplored prospect, publically available comprehensive studies of the

Canterbury Basin are somewhat limited. The modern seafloor canyons shown in Figure 9 are good

analogs of the deeper paleocanyons illuminated by the 3D seismic amplitude and attribute data.

ATTRIBUTE SELECTION

In their comparison of alternative unsupervised learning techniques, Barnes and Laughlin

(2002) concluded that the appropriate choice of attributes was the most critical component of

computer assisted seismic facies identification. Although interpreters are skilled at identifying

facies, such recognition is often subconscious and hard to define (see Eagleman’s 2012 discussion

on differentiating male from female chicks and identifying military aircraft from silhouettes). In

supervised learning, the software does some of the work during the training process, though we

must always be wary of false correlations if we provide too many attributes (Kalkomey, 1999).

For the prediction of continuous data such as porosity, Russell (1997) and others suggest that one

begin with exploratory data analysis, where one simply cross-correlates a candidate attribute with

the desired property at the well. Such cross-correlation does not work well when trying to identify

seismic facies, which are simply “labeled” with an integer number or alphanumeric name.

Table 1 summarizes how we four interpreters perceive each of the seismic facies of interest.

Once we have enumerated the seismic expression, the quantification using attribute expression is

relatively straightforward. In general, amplitude and frequency attributes are lithology indicators

and may provide direct hydrocarbon detection in conventional reservoirs, geometric attributes

delineate reflector morphology such as dip, curvature, rotation, and convergence, while statistical

and texture attributes provides information about data distribution that quantifies subtle patterns

19

that are hard to define (Chopra and Marfurt, 2007). Attributes such as coherence provide images

of the edges of seismic facies rather than a measure of the facies themselves, although slumps often

appear as a suite of closely spaced faults separating rotated fault blocks. Finally, what we see as

interpreters and what our clustering algorithms see can be quite different. While we may see a

slump feature as exhibiting a high number of faults per km, our clustering algorithms are applied

voxel by voxel and see only the local behavior. Extending the clustering to see such large scale

textures requires the development of new texture attributes.

The number of attributes should be as small as possible to discriminate the facies of interest,

and each attribute should be mathematical independent from the others. While it may be fairly

easy to represent three attributes with a deformed 2D manifold, increasing the dimensionality

results in increased deformation, such that our manifold may fold on itself or may not accurately

represent the increased data variability. Because the Waka-3D survey is new to all four authors,

we have tested numerous attributes that we think may highlight different facies in the turbidite

system. Among these attributes, we find the shape index to be good for visual classification but

dominates the unsupervised classifications with valley and ridge features across the survey. After

such analysis we chose four attributes that are mathematically independent but should be coupled

through the underlying geology: peak spectral frequency, peak spectral magnitude, GLCM

homogeneity, and curvedness, as the input to our classifiers. The peak spectral frequency and peak

spectral magnitude form an attribute pair that crudely represents the spectral response. Peak

frequency of spectrally whitened data is sensitive to tuning thickness while peak magnitude is a

function of both tuning thickness and impedance contrast. GLCM homogeneity is a texture

attribute that has a high value for adjacent traces with similar (high or low) amplitudes and

measures the continuity of a seismic facies. Curvedness defines the magnitude of reflector

20

structural or stratigraphic deformation, with dome-, ridge-, saddle-, valley-, and bowl-shaped

features exhibiting high curvedness and planar features exhibiting zero curvedness.

Figure 10 shows a time slice at t=1.88 s through the seismic amplitude volume on which

we identify channels (white arrows), high amplitude deposits (yellow arrows), and slope fans (red

arrows). Figure 11 shows an equivalent time slice through peak spectral frequency co-rendered

with peak spectral magnitude that emphasizes the relative thickness and reflectivity of the turbidite

system and surrounding slope fan sediments into which it was incised. The edges of the channels

are delineated by Sobel filter similarity. We show equivalent time slices through (Figure 12)

GLCM homogeneity, and (Figure 13) co-rendered shape index and curvedness. In Figure 14 we

show a representative vertical slice at line AA’ in Figure 10 cutting through the channels through

(Figure 14a) seismic amplitude, (Figure 14b) seismic amplitude co-rendered with peak spectral

magnitude/peak spectral frequency, (Figure 14c) seismic amplitude co-rendered with GLCM

homogeneity, and (Figure 14d) seismic amplitude co-rendered shape index and curvedness. White

arrows indicate incised valleys, yellow arrows high amplitude deposits, and red arrows a slope fan.

We note several of the incised values are visible at time slice t=1.88 s.

In a conventional interpretation workflow, the geoscientist would examine each of these

attribute images and integrate them within a depositional framework. Such interpretation takes

time and may be impractical for extremely large data volumes. In contrast, in seismic facies

classification the computer either attempts to classify what it sees as distinct seismic facies (in

unsupervised learning) or attempts to emulate the interpreter’s classification made on a finite

number of vertical sections, time, and/or horizon slices and apply the same classification to the full

3D volume (in supervised learning). In both cases, the interpreter needs to validate the final

classification to determine if they represent seismic facies of interest. In our example we will use

21

Sobel filter similarity to separate the facies and then evaluate how they fit within our understanding

of a turbidite system.

APPLICATION

Given these four attributes, we now construct four-dimensional attribute vectors as input

to the previously described classification algorithms. To better illustrate the performance of each

algorithm, we summarize the data size, number of computational processors, and runtime in Table

2. All the algorithms are developed by the authors except ANN, which is implemented using

MATLAB® toolbox.

We begin with k-means. As previously discussed, a limitation of k-means is the lack of any

structure to the cluster number selection process. We illustrate this limitation by computing k-

means with 16 (Figure 15) and 256 (Figure 16) clusters. On Figure 15, we can identify high

amplitude overbank deposits (yellow arrows), channels (white arrows), and slope fan deposits (red

arrows). A main limitation of k-means is that there is no structure linking the clusters, which leads

to a somewhat random choice of color assignment to clusters. This problem becomes more serious

when more clusters are selected: the result with 256 clusters (Figure 16) is so chaotic that we can

rarely separate the overbank high amplitude deposits (yellow arrows) and slope fan deposits (red

arrows) that were easily separable in Figure 15. For this reason, modern k-means applications focus

on estimating the correct number of clusters in the data.

In contrast to k-means, SOM restricts the cluster centers to lie on a deformed 2D manifold.

While clusters may move closer or further apart, they still form (in our implementation) a deformed

quadrilateral mesh which maps to a rectangular mesh on the 2D latent space. Mapping the latent

space to a continuous 1D (Coleou et al., 2003) or 2D color bar (Strecker and Uden, 2002), reduces

22

the sensitivity to the number of clusters chosen. We follow Gao (2007) and avoid guessing at the

number of clusters necessary to represent the data by overdefining the number of prototype vectors

to be 256 (the limit of color levels in our commercial display software). These 256 prototype

vectors (potential clusters) reduce to only three or four distinct “natural” clusters through the SOM

neighborhood training criteria. The 2D SOM manifold is initialized using the first two principle

components, defining a plane through the N-dimensional attribute space (Figure 17). The

algorithm then deforms the manifold to better fit the data. Overdefining the number of prototype

vectors results in clumping into a smaller number natural clusters. These clumped prototype

vectors project onto adjacent locations in the latent space are therefore appear as subtle shades of

the same color as indicated by the limited palette of 256 colors shown in Figure 18. On the

classification result shown on Figure 18, we can clearly identify the green colored spill-over

deposits (yellow arrows). The difference between channel fill (white arrows) and slope fans (red

arrows) is insignificant. However, by co-rendering with similarity, the channels are delineated

nicely, allowing us to visually distinguish channel fills and the surrounded slope fans. We can

also identify some purple color clusters (orange arrows) which we interpret to be crevasse splays.

Next, we apply GTM to the same four attributes. We compute two “orthogonal” projections

of data onto the manifold and thence onto the two dimensions of the latent space (Figure A2).

Rather than define explicit clusters, we project the mean a posteriori probability distribution onto

the 2D latent space and then export the projection onto the two latent space axes. We crossplot the

projections along axes 1 and 2 and map them against a 2D color bar (Figure 19). In this slice, we

see channels delineated by purple colors (white arrows), point bar and crevasse splays in pinkish

colors (yellow arrows), and slope fans in lime green colors (red arrows). We can also identify some

thin, braided channels at the south end of the survey (blue arrow). Similarly to the SOM result,

23

similarity separates the incised valleys from the slope fans. However, the geological meaning of

the orange colored facies is somehow vague. This is the nature of unsupervised learning techniques

in that the clusters represent topological differences in the input data vectors, which are not

necessarily the facies differences we wish to delineate. We can ameliorate this shortcoming by

adding a posteriori supervision to the GTM manifold. The simplest way to add supervision is to

compute the average attribute vectors about a given seismic facies and map it to the GTM

crossplot. Then, the interpreter can manually define clusters on the 2D histogram by constructing

one or more polygons (Figure 20), where we cluster the data into four facies: multistoried channels

(blue), high-energy point bar and crevasse splay deposits (yellow), slope fans (green), and

“everything else” (red). A more quantitative methodology is to mathematically project these

average clusters onto the manifold, and then cross multiply the probability distribution of the

control vectors against the probability distribution function of each data vector, thereby forming

the Bhattacharya distance (Roy et al., 2013, 2014). Such measures then provide a probability

ranging between 0 and 100% as to whether the data vector at any seismic sample point is like the

data vectors about well control (Roy et al., 2013, 2014) or like the average data vector within a

facies picked by the interpreter.

The a posteriori supervision added to GTM is the critical prior supervision necessary for

supervised classification such as ANN (Figure A3) and SVM (Figure A4). In this study we used

the same four attributes as input for both unsupervised and supervised learning techniques. Our

supervision consists of picked seed points for the three main facies previously delineated using the

unsupervised classification results, which are multistoried channel, point bar and crevasse splay

deposits, and slope fans, plus an additional channel flank facies. The seed points are shown in

Figure 21. Seed points should be picked with great caution to correctly represent the corresponding

24

facies, any false picking (a seed point that does not belong to the intended facies) will greatly

compromise the classification result. We then compute averages of the four input attributes within

a 7 trace X 7 trace X 24 ms window about each seed point to generate a training table which consists

of 4-dimensional input attribute vectors and one dimensional targets (the labeled facies).

For our ANN application, we used the neural networks toolbox in MATLAB®, and

generated a probabilistic neural network (PNN) composed of 20 neurons. Because of the relatively

small size of the training data, the training process only took a second or so; however, since a PNN

may converge to local minima, we are not confident that our first trained network has the best

performance. Our workflow is then to rerun the training process 50 times and choose the network

exhibiting the lowest training and cross-validation errors. Figures 22 and 23 show the PNN

performance during training, while Figure 24 shows the PNN classification result. We notice that

all the training, testing, and cross-validation performance are acceptable, with training and cross-

validation correctness being around 90%, and testing correctness being over 86%. We identify

blue channel stories within the relatively larger scale incised valleys (white arrows), and yellow

point bars and crevasse splays (yellow arrows). However, many of the slope fan deposits are now

classified as channel flanks or multistoried channels (blue arrows), which need to be further

calibrated with well log data. Nevertheless, as a supervised learning technique, ANN provides

classification with explicit geological meaning, which is its primary advantage over unsupervised

learning techniques.

Finally, we cluster our four-dimensional input data using SVM, using the same training

data (interpreter picks) as for ANN. The workflow is similar to ANN in that we ran 20 passes of

training, varying the Gaussian kernel standard deviation, σ, and misclassification tolerance, ε,

parameters for each pass. These parameter choices are easier than selecting the number of neurons

25

for ANN, since the SVM algorithm solves a convex optimization problem that converges to a

global minima. The training and cross-validation performance is comparable to ANN, with

roughly 92% training correctness and 85% cross-validation correctness. Figure 25 shows the SVM

classification result at time t = 1.88 s. The SVM map follows the same pattern as we have seen

on the ANN map, but is generally cleaner, with some differences in details. Compared to ANN,

SVM successfully mapped more of the slope fans (white arrows), but missed some crevasse splays

that were correctly picked by ANN (yellow arrow). We also see a great amount of facies variation

within the incised valleys, which is reasonable because of the multiple course changes of a

paleochannel during its deposition that results in multiple channel stories. Finally, we note some

red lines following NW-SE direction (red arrows) which correspond to acquisition footprint.

CONCLUSIONS

In this paper we have compared and contrasted some of the more important multiattribute

facies classification tools, including four unsupervised (PCA, k-means, SOM, GTM) and two

supervised (ANN, SVM) learning techniques. In addition to highlighting the differences in

assumptions and implementation, we have applied each method to the same Canterbury Basin

survey, with the goal of delineating seismic facies in a turbidite system to demonstrate the

effectiveness and weaknesses of each method. K-means and SOM move the user-defined number

of cluster centers towards the input data vectors. PCA is the simplest manifold method, where the

data variability in our examples is approximated by a 2D plane defined by the first two

eigenvectors. GTM is more accurately described as a mapping technique, like PCA, where the

clusters are formed either in the human brain as part of visualization or through crossplotting and

the construction of polygons. SOM and GTM manifolds deform to fit the N-dimensional data. In

SOM, the cluster centers (prototype vectors) move along the manifold towards the data vectors,

26

forming true clusters. In all four methods, any labeling of a given cluster to a given facies happens

after the process is completed. In contrast, ANN and SVM build a specific relation between the

input data vectors and a subset of user-labeled input training data vectors thereby explicitly

labeling the output clusters to the desired facies. Supervised learning is constructed from a limited

group of training samples (usually at certain well locations or manually picked seed points) which

generally are insufficient to represent all the lithologic and stratigraphic variations within a

relatively large seismic data volume. A pitfall of supervised learning is that unforeseen clusters

will be misclassified as clusters that have been chosen.

For this reason, unsupervised classification products can be used to construct not only an

initial estimate of the number of classes, but also a validation tool to determine if separate clusters

have been incorrectly lumped together. We advise computing unsupervised SOM or GTM prior to

picking seed points for subsequent supervised learning, to clarify the topological differences

mapped by our choice of attributes. Such mapping will greatly improve the picking confidence,

because the seed points are now confirmed by both human experience and mathematical statistics.

The choice of the correct suite of attributes is critical. Specifically, images that are ideal

for multiattribute visualization may be suboptimal for clustering. We made several poor choices

in previous iterations of writing this paper. The image of inline (SW-NE) structural dip illustrates

this problem directly. While a skilled interpreter sees a great deal of detail in Figure 26, there is no

clear facies difference between positive and negative dips, such that this component of vector dip

cannot be used to differentiate them. A better choice would be dip magnitude, except that a long

wavelength overprint (such as descending into the basin) would again bias our clustering in a

manner that is unrelated to facies. Therefore, we tried to use relative changes in dip – curvedness

27

and shape indices measure lateral changes in dip, and reflector convergence which differentiates

conformal from nonconformal reflectors.

Certain attributes should never be used in clustering. Phase, azimuth, and strike have

circular distributions, where a phase value of -180 indicates the same value as +180. No trend can

be found. While the shape index, s, is not circular, ranging between -1 and +1, the histogram has

a peaks about the ridge (s=+0.5) and about the valley (s=-0.5). We speculate that shape components

may be more amenable to classification. Reflector convergence follows the same pattern as

curvedness. For this reason we only used curvedness as a representative of these three attributes.

The addition of this choice improved our clustering.

Edge attributes like the Sobel filter similarity and coherence are not useful for the example

show here; instead, we have visually added them as an edge “cluster” and co-rendered with the

images shown in Figures 15-21, 24, and 25. In contrast, when analyzing more chaotic features

such as salt domes and karst collapse, coherence is a good input to clustering algorithms. We do

wish to provide an estimate of continuity and randomness to our clustering. To do so, we follow

Corradi et al. (2009) and West et al. (2002) and use GLCM homogeneity as an input attribute.

Theoretically, no one technique is superior to all the others in every aspect, and each

technique has its inherent advantages and defects. K-means with a relatively small numbers of

clusters is the easiest algorithm to implement, provides rapid interpretation, but lacks the relation

among clusters. SOM provides a generally more “interpreter friendly” clustering result with

topological connections among clusters, but is computationally more demanding than k-means.

GTM relies on probability theory and enables the interpreter to add posteriori supervision by

manipulating the data’s posterior probability distribution; however, it is not widely accessible to

28

the exploration geophysicist community. Rather than displaying the conventional cluster numbers

(or labels), we suggest displaying the cluster coordinates projected onto the 2D SOM and GTM

latent space axes. Doing so not only provides greater flexibility in constructing a 2D color bar but

also provides data that can be further manipulated using 2D crossplot tools.

For the two supervised learning techniques, ANN suffers from the convergence problem

and requires expertise to achieve the optimal performance, while the computation cost is relatively

low. SVM is mathematically more robust and easier to train, but is more computationally

demanding.

Practically, if no software limitations are set, we can make suggestions on how an

interpreter can incorporate these techniques to facilitate seismic facies interpretation at different

exploration and development stages. To identify the main features in a recently acquired 3D

seismic survey on which limited to no traditional structural interpretation is done, k-means is a

good candidate for exploratory classification starting with a small K (typically K = 4) and gradually

increase the number of class. As more data are acquired (e.g. well log data and production data)

and detailed structural interpretation has been performed, SOM or GTM focusing in the target

formations will provide more refined classification, which needs to be calibrated with wells. In the

development stage when most of the data have been acquired, with proper training process, ANN

and SVM provide targeted products, characterizing the reservoir by mimicking interpreters’

behavior. Generally, SVM provides superior classification than ANN but at a considerably higher

computational cost, so choosing between these two requires balancing performance and runtime

cost. As a practical manner, no given interpretation software platform provides all five of these

clustering techniques, such that many of the choices are based on software availability.

29

Because we wish this paper to serve as an inspiration of interpreters, we do want to reveal

one drawback of our work: all the classifications are performed volumetrically but not along a

certain formation. Such classification may be biased by the bonding formations above and below

the target formation (if we do have a target formation), therefore contaminates the facies map.

However, we want to make the point that such classification can happen at a very early stage of

interpretation, when both structural interpretation and well logs are very limited. And even in such

situation, we can still use classification techniques to generate facies volumes to assist subsequent

interpretation.

In the 1970s and 1980s much of geophysical innovation in seismic processing and

interpretation was facilitated by the rapid evolution of computer technology – from mainframes to

minicomputers to workstations to distributed processing. We believe similar advances in facies

analysis will be facilitated by the rapid innovation in “big data” analysis, driven by needs in

marketing and security. While we may not answer Turing’s (1950) question “Can machines

think?”, we will certainly be able to teach them how to emulate a skilled human interpreter.

ACKNOWLEDGEMENTS

We thank New Zealand Petroleum and Minerals for providing the Waka-3D seismic data

to the public. Financial support for this effort and for the development of our own k-means, PCA,

SOM, GTM, and SVM algorithms was provided by the industry sponsors of the Attribute-Assisted

Seismic Processing and Interpretation (AASPI) consortium in the University of Oklahoma. The

ANN examples were run using MATLAB®. We thank colleagues Fangyu Li, Sumit Verma,

Lennon Infante, and Brad Wallet for their valuable input and suggestions. All the 3D seismic

displays were made using licenses to Petrel, provided to the University of Oklahoma for research

30

and education courtesy of Schlumberger. Finally, we want to express our great respect to the

people that have contributed the development of pattern recognition techniques in exploration

geophysics field.

31

APPENDIX: Mathematical details

In this appendix we summarize many of the mathematical details defining the various

algorithm implementations. Although insufficient to allow a straightforward implementation of

each algorithm, we hope to more quantitatively illustrate the algorithmic assumptions as well as

algorithmic similarities and differences. Because k-means and artificial neural networks have been

widely studied, in this appendix we only give some principle statistical background, and brief

reviews of SOM, GTM, and PSVM algorithms involved in this tutorial. We begin this appendix

by giving statistical formulations of the covariance matrix, principal components and the

Mahalanobis distance when applied to seismic attributes. We further illustrate the formulations

and some necessary theory for SOM, GTM, ANN, and PSVM. Because of the extensive use of

mathematical symbols and notations, a table of shared mathematical notations is given in Table

A1. All other symbols are defined in the text.

Covariance matrix, principal components, and the Mahalanobis distance

Given a suite of N attributes, the covariance matrix is defined as

𝐶𝑚𝑛 =1

𝐽∑(𝑎𝑗𝑚(𝑡𝑗 , 𝑥𝑗 , 𝑦𝑗) − 𝜇𝑚)(𝑎𝑗𝑛(𝑡𝑗 , 𝑥𝑗 , 𝑦𝑗) − 𝜇𝑛)

𝐽

𝑗=1

, (𝐴 − 1)

where ajm and ajn are the mth and nth attributes, J is the total number of data vectors, and where

𝜇𝑛 =1

𝐽∑ 𝑎𝑗𝑛(𝑡𝑗 , 𝑥𝑗 , 𝑦𝑗)

𝐽

𝑗=1

, (𝐴 − 2)

is the mean of the nth attribute. If we compute the eigenvalues, λi, and eigenvectors, vi, of the real,

symmetric covariance matrix, C, the ith principal component at data vector j is defined as

𝑝𝑗𝑖 = ∑ 𝑎𝑗𝑛(𝑡𝑗 , 𝑥𝑗 , 𝑦𝑗)

𝑁

𝑛=1

𝑣𝑛𝑖 , (𝐴 − 3)

32

where vni indicates the nth attribute component of the ith eigenvector. In this paper, the first two

eigenvectors and eigenvalues are also used to construct an initial model in both the self-organizing

maps (SOM) and generative topological mapping (GTM) algorithms.

The Mahalanobis distance, rjq, of the jth sample from the qth cluster center, θq, is defined as

𝑟𝑗𝑞2 = ∑ ∑ (𝑎𝑗𝑛 − 𝜃𝑛𝑞)𝐶𝑛𝑚

−1 (𝑎𝑗𝑚 − 𝜃𝑚𝑞)

𝑁

𝑚=1

𝑁

𝑛=1

, (𝐴 − 4)

where the inversion of the covariance matrix, C, takes place prior to extracting the mnth element.

Self-organizing maps

Rather than computing the Mahalanobis distance, both SOM and GTM first normalize the

data using a Z-scale. If the data exhibit an approximately Gaussian distribution, the Z-scale of the

nth attribute is obtained by subtracting the mean and dividing by the standard deviation (the square

root of the diagonal of the covariance matrix, Cnn). To Z-scale non-Gaussian distributed data, such

as coherence, one needs to first break the data using histograms that approximate a Gaussian. The

objective of the SOM algorithm is to map the input seismic attributes onto a geometric manifold

called the self-organized map. The SOM manifold is defined by a suite of prototype vectors mk

lying on a lower-dimensional (in our case, 2D) surface which fit the N-dimensional attribute data.

The prototype vectors mk are typically arranged in 2D hexagonal or rectangular structure maps

that preserve their original neighborhood relationship, such that neighboring prototype vectors

represent similar data vectors. The number of prototype vectors in the 2D map determines the

effectiveness and generalization of the algorithm. One strategy is to estimate the number of initial

clusters, and then to either divide or join clusters based on distance criteria. In our case, we follow

Gao (2007) and overdefine the number of clusters to be the maximum number of colors supported

33

by our visualization software. Interpreters then either use their color perception or construct

polygons on 2D histograms to define a smaller number of clusters.

Our implementation of the SOM algorithm is summarized in Figure A1. After computing

Z-scores of the input data, the initial manifold is defined to be a plane defined by the two first

principal components. Prototype vectors mk are defined on a rectangular grid to the first two

eigenvalues to range between 2(λ1)1/2 and 2(λ2)

1/2. The seismic attribute data are then compared

to each of the prototype vectors, finding the nearest one. This prototype vector and its nearest

neighbors (those that fall within a range σ, defining a Gaussian perturbation) are moved towards

the data point. After all the training vectors have been examined, the neighborhood radius, σ, is

reduced. Iterations continue until σ approaches the distance between the original prototype vectors.

Given this background, Kohonen (2001) defines the SOM training algorithm using the following

five steps:

Step 1: Randomly chose a previously Z-scored input attribute vector, 𝐚𝑗, from the set of

input vectors.

Step 2: Compute the Euclidean distance between this vector 𝐚𝑗 and all prototype

vectors 𝐦𝑘 , k=1,2,…,K. The prototype vector which has the minimum distance to the input

vector 𝐚𝑗, is defined to be the “winner” or the Best Matching Unit, 𝐦𝑏:

||𝐚𝑗 − 𝐦𝑏|| = 𝑚𝑖𝑛𝑘

{||𝐚𝑗 − 𝐦𝑘||}. (𝐴 − 5)

Step 3: Update the “winner” prototype vector and its neighbors. The updating rule for the

weight of the kth prototype vector inside and outside the neighborhood radius 𝜎(𝑡) is given by

𝐦𝑘(𝑡 + 1) = {𝐦𝑘(𝑡) + 𝛼(𝑡)ℎ𝑏𝑘(𝑡)[𝐚𝑗 − 𝐦𝑘(𝑡)], 𝑖𝑓 ||𝐫𝑘 − 𝐫𝑏|| ≤ 𝜎(𝑡)

𝐦𝑘(𝑡), 𝑖𝑓 ||𝐫𝑘 − 𝐫𝑏|| > 𝜎(𝑡) (𝐴 − 6)

34

where the neighborhood radius defined as 𝜎(𝑡) is predefined for a problem and decreases with

each iteration 𝑡. 𝐫𝑏 and 𝐫𝑘 are the position vectors of the winner prototype vector 𝐦𝑏 and the

kth prototype vector 𝐦𝑘 respectively. We also define the neighborhood function, ℎ𝑏𝑘(𝑡), the

exponential learning function, 𝛼(𝑡), and the length of training, T. ℎ𝑏𝑘(𝑡) and 𝛼(𝑡) decrease with

each iteration in the learning process and are defined as

ℎ𝑏𝑘(𝑡) = 𝑒−(||𝐫𝑏−𝐫𝑘||2/2𝜎2(𝑡), (𝐴 − 7)

and

𝛼(𝑡) = 𝛼0(0.005

𝛼0)𝑡/𝑇 . (𝐴 − 8)

Step 4: Iterate through each learning step (steps 1-3) until the convergence criterion (which

depends on the predefined lowest neighborhood radius and the minimum distance between the

prototype vectors in the latent space) is reached.

Step 5: Project the prototype vectors onto the first two principal components and color code

using a 2D color bar (Matos et al. 2009).

Generative topological mapping

In GTM, the grid points of our 2D deformed manifold in N-dimensional attribute space

define the centers, mk, of Gaussian distributions of variance σ2=β-1. These centers, mk, are in turn

projected onto a 2D latent space, defined by a grid of nodes uk and nonlinear basis functions, Φ:

𝐦𝑘 = ∑ 𝐖𝑘𝑚𝛷𝑚(𝐮𝑘)

𝑀

𝑚=1

, (𝐴 − 11)

where W is a K X M matrix of unknown weights, 𝛷𝑚(𝐮𝑘) is a set of M nonlinear basis functions,

mk are vectors defining the deformed manifold in the N-dimensional data space, and k=1,2,…,K

is the number of grid points arranged on a lower L-dimensional latent space (in our case, L=2). A

noise model (the probability of the existence of a particular data vector aj given weights W and

35

inverse variance β) is introduced for each measured data vector. The probability density function,

p, is represented by a suite of K radially symmetric N-dimensional Gaussian functions centered

about mk with variance of 1/β:

𝑝(𝐚𝑗|𝐖, 𝛽) = ∑1

𝐾

𝐾

𝑘=1

(𝛽

2𝜋)

𝑁2 𝑒−

𝛽2

||𝐦𝑘−𝐚𝑗||2

. (𝐴 − 12)

The prior probabilities of each of these components are assumed to be equal with a value of 1/K,

for all data vectors 𝐚𝑗. Figure 4 illustrates the GTM mapping from an L=2D latent space to the 3D

data space.

The probability density model (GTM model) is fit to the data 𝐚𝑗 to find the parameters W

and β using a maximum likelihood estimation. One of the popular techniques used in parameter

estimations is the Expectation Maximization (EM) algorithm. Using Bayes’ theorem, and the

current values of the GTM model parameters W and β, we calculate the J X K posterior probability

or responsibility, Rjk, for each of the K components in latent space for each data-vector:

𝑅𝑗𝑘 =𝑒

−𝛽2

||𝐦𝑘−𝐚𝑗||2

∑ 𝑒−

𝛽2

||𝐦𝑖−𝐚𝑗||2

𝑖

. (𝐴 − 13)

Equation A-13 forms the “E-step” or Expectation step in the EM algorithm. The E-step is

followed by the Maximization or “M-step”, which uses these responsibilities to update the model

for a new weight matrix W by solving a set of linear equations (Dempster et al., 1977):

(𝚽𝑇𝐆𝚽 + 𝛼

𝛽𝐈) 𝐖𝑇

𝑛𝑒𝑤 = 𝚽𝑇𝐑𝐗, (𝐴 − 14)

where

𝐺𝑘𝑘 = ∑ 𝑅𝑗𝑘𝐽𝑗=1 are the non-zero elements of the K X K diagonal matrix G,

𝚽 is a K X M Matrix with elements 𝚽 = 𝛷𝑚(𝐮𝑘),

36

𝛼 is a regularization constant to avoid division by zero, and

I is the M X M identity matrix.

The updated value of β is given by

1

𝛽𝑛𝑒𝑤=

1

𝐽𝑁∑ ∑ 𝑅𝑗𝑘

𝐾

𝑘=1

𝐽

𝑗=1

||𝐖𝑘𝑚𝑛𝑒𝑤𝛷𝑚(𝐮𝑘) − 𝐚𝑗||

2

. (𝐴 − 15)

The initialization of W is done so that the initial GTM model approximates the principal

components (largest eigenvectors) of the input data, 𝐚𝑗. The value of β-1 is initialized to be the

larger of the (L+1)th eigenvalue from PCA where L is the dimension of the latent space. In Figure

4, L=2, such that we initialize β-1 to be the inverse of the third eigenvalue. Figure A2 summarizes

this workflow.

Artificial Neural Networks

The artificial neural networks are a class of pattern recognition algorithms that were

derived separately in different fields such as statistics and artificial intelligence. Artificial neural

networks are easily accessible for most of the geophysical interpreters, so we only provide a

general workflow of applying an ANN to seismic facies classification for completeness of this

tutorial. The workflow is shown in Figure A3.

Proximal support vector machines

Because SVMs are originally developed to solve binary classification problems, the

arithmetic we begin with a summary of the arithmetic describing a binary PSVM classifier.

Similarly to SVM, a PSVM decision condition is defined as (Figure 7):

𝐱𝑇𝛚 − 𝛾 {> 0, 𝐱 ∈ 𝑋+;= 0, 𝐱 ∈ 𝑋 + 𝑜𝑟 𝑋−;< 0, 𝐱 ∈ 𝑋−,

(𝐴 − 16)

37

where 𝐱 is an N-dimensional attribute vector to be classified, 𝛚 is a N X 1 vector implicitly defines

the normal of the decision-boundary in the higher dimensional space, 𝛾 defines the location of the

decision-boundary, and “𝑋 +” and “𝑋 −” indicate the two classes of the binary classification.

PSVM solves an optimization problem and takes the form of (Fung and Mangasarian, 2001):

𝑚𝑖𝑛𝛚,𝛾,𝐲

𝜀1

2‖𝐲‖2 +

1

2(𝛚𝑇𝛚 + 𝛾2), (𝐴 − 17)

subject to

𝐃(𝐚𝛚 − 𝐞𝛾) + 𝐲 = 𝐞. (𝐴 − 18)

In this optimization problem, 𝐲 is a J X 1 error variable; 𝐚 is a J X N sample matrix

composed of 𝐽attribute vectors, which can be divided into two classes, “𝑋 +” and “𝑋 −”. 𝐃 is a J

X J diagonal matrix of labels with a diagonal composed of “+1” for “𝑋 +” and “−1” for “𝑋 −”. 𝜀

is a non-negative parameter. Finally, 𝐞 is a J X 1 column vector of ones. This optimization problem

can be solved by using a J X 1 Lagrangian multiplier 𝐭:

𝐿(𝛚, 𝛾, 𝐲, 𝐭) = 𝜀1

2‖𝐲‖2 +

1

2(𝛚𝑇𝛚 + 𝛾2) − 𝐭𝑇(𝐃(𝐚𝛚 − 𝐞𝛾) + 𝐲 − 𝐞). (𝐴 − 19)

By setting the gradients of 𝐿 to zero, we obtain expressions for 𝛚, 𝛾 and 𝐲 explicitly in the

knowns and 𝐭 , where 𝐭 can further be represented by 𝐚 , 𝐃 and 𝜀 . Then by substituting 𝛚 in

Equations A-17 and A-18 using its dual equivalent 𝛚 = 𝐚𝑇𝐃𝐭 , we can arrive at (Fung and

Mangasarian, 2001):

𝑚𝑖𝑛𝛚,𝛾,𝐲

𝜀1

2‖𝐲‖2 +

1

2(𝐭𝑇𝐭 + 𝛾2), (𝐴 − 20)

subject to

𝐃(𝐚𝐚𝑇𝐃𝐭 − 𝐞𝛾) + 𝐲 = 𝐞. (𝐴 − 21)

38

Equations A-20 and A-21 provide a more desirable version of the optimization problem

since one can now insert kernel methods to solve nonlinear classification problems made possible

by the term 𝐚𝐚𝑇 in Equation A-21. Utilizing the Lagrangian multiplier again (this time we denote

the multiplier as 𝛕), we can minimize the new optimization problem against 𝐭, 𝛾, 𝐲 and 𝛕. By

setting the gradients of these four variables to zero, we can express 𝐭, 𝛾 and 𝐲 explicitly by 𝛕 and

other knowns, where 𝛕 is solely a dependent on the data matrices. Then for N-dimensional attribute

vector 𝐱 we write the decision conditions as

𝐱𝑇𝐚𝑇𝐃𝐭 − 𝛾 {> 0, 𝐱 ∈ 𝑋+;= 0, 𝐱 ∈ 𝑋 + 𝑜𝑟 𝑋−;< 0, 𝐱 ∈ 𝑋−,

(𝐴 − 22)

with

𝐭 = 𝐃𝐊𝑇𝐃 (𝐈

𝜀+ 𝐆𝐆𝑇)

−1

𝐞, (𝐴 − 23)

𝛾 = 𝐞𝑇𝐃 (𝐈

𝜀+ 𝐆𝐆𝑇)

−1

𝐞, (𝐴 − 24)

and

𝐆 = 𝐃[𝐊 −𝐞]. (𝐴 − 25)

Instead of 𝐚, we have 𝐊 in Equations A-23 and A-25, which is a Gaussian kernel function

of 𝐚 and 𝐚𝑇 that has the form of:

𝐊(𝐚, 𝐚𝑇)𝑖𝑗 = 𝑒𝑥𝑝 (−𝜎‖𝐚𝑇𝑖∙ − 𝐚𝑇

𝑗∙‖2

) , 𝑖, 𝑗 ∈ [1, 𝐽], (𝐴 − 26)

where 𝜎 is a scalar parameter. Finally, by replacing 𝐱𝑇𝐚𝑇 by its corresponding kernel

expression, the decision condition can be written as:

𝐊(𝐱𝑇, 𝐚𝑇)𝐃𝐭 − 𝛾 {> 0, 𝐱 ∈ 𝑋+;= 0, 𝐱 ∈ 𝑋 + 𝑜𝑟 𝑋−;< 0, 𝐱 ∈ 𝑋−.

(𝐴 − 27)

and

39

𝐊(𝐱𝑇, 𝐚𝑇)𝑖𝑗 = 𝑒𝑥𝑝(−𝜎‖𝐱 − 𝐚𝑇𝑖∙‖

2) , 𝑖 ∈ [1, 𝐽]. (𝐴 − 28)

The formulations above represent a nonlinear PSVM classifier.

To extend this binary classifier to handle multiclass classification problems, some

strategies have been developed by researchers, which generally fall into three categories: “one-

versus-all”, “one-versus-one” and “all together”. For Q classes, the former two strategies build a

suite of binary classifiers individually: (𝑄(𝑄 − 1)/2 for the “one-versus-one” and 𝑄 for the “one-

versus-all” algorithm, and then use these classifiers to construct the final classification decision.

The “all together” attempts to solve multiclass problem in one step. Hsu and Lin (2002) found

“one-versus-one” method to be superior for large problems. There are two particular algorithms

for “one-versus-one” strategies, namely the “Max Wins” (Kreßel, 1999) and directed acyclic graph

(DAG) (Platt et al., 2000) algorithms. Both algorithms provide comparable results while

surpassing the “one-versus-all” method in accuracy and computational efficiency.

Our approach uses a classification factor table to assign classes to unknown samples

(Figure A4). A classification factor of an unknown sample point for a certain pilot class “X” is the

normalized distance to the binary decision boundary between “X” and the other class used when

generating this binary decision boundary. An example of a classification factor table is shown in

Figure A4, and based on this table, the unknown sample point belongs to class “D”.

40

REFERENCES

Al-Anazi, A. and I. D. Gates, 2010, A support vector machine algorithm to classify lithofacies and

model permeability in heterogeneous reservoirs: Engineering Geology, 114, 267-277.

Balch, A. H., 1971, Color sonagrams: A new dimension in seismic data interpretation: Geophysics,

36, 1074-1098.

Barnes, A. E., and K. J. Laughlin, 2002, Investigation of methods for unsupervised classification

of seismic data: 72nd Annual International Meeting, SEG, Expanded Abstracts, 2221-2224.

Bennett, K. P. and A. Demiriz, 1999, Semi-supervised support vector machines: Advances in

Neural Information Processing Systems 11: Proceedings of the 1998 Conference, 368-374.

Bishop, C. M., 2006, Pattern recognition and machine learning: Springer, New York, United

States.

Bishop, C. M., M. Svensen, and C. K. I. Williams, 1998, The generative topographic mapping:

Neural Computation, 10, 215-234.

Chao, J., M. Hoshino, T. Kitamura, and T. Masuda, 2001, A multilayer RBF network and its

supervised learning: International Joint Conference on Neural Networks, INNS/IEEE,

Expanded Abstracts, 1995–2000.

Chapelle, O., B. Schölkopf, and A. Zien, 2006, Semi-supervised learning: MIT press, Cambridge,

United States.

Chopra, S. and K. J. Marfurt, 2007, Seismic attributes for prospect identification and reservoir

characterization: Society of Exploration Geophysicists, Tulsa, United States.

41

Coleou, T., M. Poupon, and K. Azbel, 2003, Unsupervised seismic facies classification: A review

and comparison of techniques and implementation: The Leading Edge, 22, 942-953.

Corradi, A., P. Ruffo, A. Corrao, and C. Visentin, 2009, 3D hydrocarbon migration by percolation

technique in an alternative sand-shale environment described by a seismic facies

classification volume: Marine and Petroleum Geology, 26, 495-503.

Cortes, C. and V. Vapnik, 1995, Support-vector networks: Machine Learning, 20, 273-297.

Cristianini, N. and J. Shawe-Taylor, 2000, An introduction to support vector machines and other

kernel-based learning methods: Cambridge University Press, New York, United States.

Dempster, A.P., N. M. Laird, and D. B. Rubin, 1977, Maximum likelihood from incomplete data

via the EM algorithm: Journal of Royal Statistical Society, Series B, 39, 1-38.

Duda, R.O., P. E. Hart, and D. G. Stork, 2001, Pattern Classification, 2nd Edition, John Wiley &

Sons, New York, United States.

Eagleman, D., 2012, Incognito: The secret lives of the brain: Pantheon Books, New York, USA.

Forgy, E. W., 1965, Cluster analysis of multivariate data: efficiency vs interpretability of

classifications: Biometrics, 21, 768-769.

Fung, G. and O. L. Mangasarian, 2001, Proximal support vector machine classifiers: Proceedings

of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and

Data Mining, ACM 2001, 77-86.

Fung, G. M. and O. L. Mangasarian, 2005, Multicategory proximal support vector machine

classifiers: Machine Learning, 59, 77-97.

42

Gao, D., 2007, Application of three-dimensional seismic texture analysis with special reference to

deep-marine facies discrimination and interpretation: An example from offshore Angola,

West Africa: AAPG Bulletin, 91, 1665-1683.

Heggland, R., Meldahl, P., Bril, B. and de Groot, P., 1999, The chimney cube, an example of semi-

automated detection of seismic objects by directive attributes and neural networks: Part II;

interpretation, 69th Annual International Meeting, SEG, Expanded Abstracts, 935-937.

Honório, B. C. Z., A. C. Sanchetta, E. P. Leite, and A. C. Vidal, 2014, Independent component

spectral analysis: Interpretation, 2, SA21-SA29.

Hsu, C. and C. Lin, 2002, A comparison of methods for multiclass support vector machines: IEEE

Transactions on Neural Networks, 13, 415-425.

Huang, Z., J. Shimeld, M. Williamson, and J. Katsube, 1996, Permeability prediction with artificial

neural network modeling in the Ventura gas field, offshore eastern Canada: Geophysics,

61, 422–436.

Jancey, R. C., 1966, Multidimensional group analysis: Australian Journal of Botany, 14, 127-130.

Jayaram, V. and B. Usevitch, 2008, Dynamic mixing kernels in Gaussian mixture classifier for

hyperspectral classification: SPIE Optics + Photonics 2008, 70750L-70750L.

Justice, J. H., D. J. Hawkins, and D. J. Wong, 1985, Multidimensional attribute analysis and pattern

recognition for seismic interpretation: Pattern Recognition, 18, 391-407.

Kohonen, T., 1982, Self-organized formation of topologically correct feature maps: Biological

Cybernetics, 43, 59-69.

43

Kreßel, U., 1999, Pairwise classification and support vector machines: Advances in Kernel

Methods - Support Vector Learning, 255-268.

Kuzma, H. A. and J. W. Rector, 2004, Nonlinear AVO inversion using Support Vector Machines:

74th Annual International Meeting, SEG, Expanded Abstracts, 203-206.

Kuzma, H. A. and J. W. Rector, 2005, The Zoeppritz equations, information theory, and support

vector machines: 75th Annual International Meeting, SEG, Expanded Abstracts, 1701-

1704.

Kuzma, H. A. and J. W. Rector, 2007, Support Vector Machines implemented on a Graphics

Processing Unit: 77th Annual International Meeting, SEG, Expanded Abstracts, 2089-2092.

Li, J. and J. Castagna, 2004, Support Vector Machine (SVM) pattern recognition to AVO

classification: Geophysical Research Letters, 31, L02609.

Lim, J., 2005, Reservoir properties determination using fuzzy logic and neural networks from well

data in offshore Korea: Journal of Petroleum Science and Engineering, 49, 182-192.

Lubo, D., K. Marfurt, and V. Jayaram, 2014, Statistical characterization and geological correlation

of wells using automatic learning Gaussian mixture models: Unconventional Resources

Technology Conference, Extended Abstract, 2014, 774-783.

Mangasarian, O. L. and E. W. Wild, 2006, Multisurface proximal support vector machine

classification via generalized eigenvalues: IEEE Transactions on Pattern Analysis and

Machine Intelligence, 28, 69-74.

44

MacQueen, J., 1967, Some methods for classification and analysis of multivariate observations:

Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability,

Volume 1: Statistics, 281--297,

Matos, M. C., K. J. Marfurt., and P. R. S. Johann, 2009, Seismic color self-organizing maps: 11th

International Congress of the Brazilian Geophysical Society, Expanded Abstracts.

Meldahl, P., R. Heggland, B. Bril, and P. de Groot, 1999, The chimney cube, an example of semi‐

automated detection of seismic objects by directive attributes and neural networks: Part I;

Methodology: 69th Annual International Meeting, SEG, Expanded Abstracts, 931-934.

Mitchell, J. and H. L. Neil, 2012, OS20/20 Canterbury – Great South Basin TAN1209 voyage

report: National Institute of Water and Atmospheric Research Ltd (NIWA).

Murat, M. E. and A. J. Rudman, 1992, Automated first arrival picking: A neural network approach:

Geophysical Prospecting, 40, 587–604.

Nazari, S., H. A. Kuzma, and J. W. Rector, 2011, Predicting permeability from well log data and

core measurements using Support Vector Machines: 81st Annual International Meeting,

SEG, Expanded Abstracts, 2004-2007.

Platt, J. C., 1998, Sequential minimal optimization: A fast algorithm for training support vector

machines: Microsoft Research Technical Report, MSR-TR-98-14.

Platt, J. C., N. Cristianini, and J. Shawe-Taylor, 2000, Large margin DAGs for multiclass

classification: NIPS, 12, 547-553.

45

Rӧth, G. and A. Tarantola, 1994, Neural networks and inversion of seismic data: Journal of

Geophysics Research, 99, 6753–6768.

Roy, A., 2013, Latent space classification of seismic facies: PhD Dissertation, The University of

Oklahoma.

Roy, A., B.L. Dowdell, and K.J. Marfurt, 2013, Characterizing a Mississippian tripolitic chert

reservoir using 3D unsupervised and supervised multiattribute seismic facies analysis: An

example from Osage County, Oklahoma: Interpretation, 1, SB109-SB124.

Roy, A., Araceli, S. R, J. T. Kwiatkowski, and K. J. Marfurt, 2014, Generative Topographic

Mapping for seismic facies estimation of a carbonate wash, Veracruz Basin, Southern

Mexico: Interpretaion, 2, SA31-SA47.

Sammon, W. J., 1969, A nonlinear mapping for data structure analysis, IEEE Transaction on

Computers, C-18, 401-409.

Schölkopf, B. and A. J. Smola, 2002, Learning with kernels: Support vector machines,

regularization, optimization, and beyond: MIT Press, Cambridge, United States.

Shawe-Taylor, J. and N. Cristianini, 2004, Kernel methods for pattern analysis: Cambridge

University Press, New York, United States.

Slatt, M. R., and Y. Abousleiman, 2011, Merging sequence stratigraphy and geomechanics for

unconventional gas shales: The Leading Edge, 30, 274-282.

Sonneland, L., 1983, Computer aided interpretation of seismic data: 53rd Annual International

Meeting, SEG, Expanded Abstracts, 546-549.

46

Strecker, U., and R. Uden, 2002, Data mining of 3D post- stack attribute volumes using Kohonen

self-organizing maps: The Leading Edge, 21, 1032-1037.

Taner, M. T., F. Koehler, and R. E. Sheriff, 1979, Complex seismic trace analysis: Geophysics,

44, 1041-1063.

Tong, S. and D. Koller, 2002, Support vector machine active learning with applications to text

classification: The Journal of Machine Learning Research, 2, 45-66.

Torres, A. and J. Reveron, 2013, Lithofacies discrimination using support vector machines, rock

physics and simultaneous seismic inversion in clastic reservoirs in the Orinoco Oil Belt,

Venezuela: 83rd Annual International Meeting, SEG, Expanded Abstracts, 2578-2582.

Turing, A.M., 1950, Computing machinery and intelligence. Mind, 59, 433-460.

Uruski, C. I., 2010, New Zealand’s deepwater frontier: Marine and Petroleum Geology, 27, 2005-

2026.

van der Baan, M. and C. Jutten, 2000, Neural networks in geophysical applications: Geophysics,

65, 1032-1047.

Vapnik, V., 1998, Statistical learning theory: Wiley & Sons, Hoboken, United States.

Verma, S., A. Roy, R. Perez, and K. J. Marfurt, 2012, Mapping high frackability and high TOC

zones in the Barnett Shale: Supervised Probabilistic Neural Network vs. unsupervised

multi-attribute Kohonen SOM: 82nd Annual International Meeting, SEG, Expanded

Abstracts: 1-5.

47

Wallet, C. B., M. C. Matos, and J. T. Kwiatkowski, 2009, Latent space modeling of seismic data:

An overview: The Leading Edge, 28, 1454-1459.

Wang, G., T. R. Carr, Y. Ju, and C. Li, 2014, Identifying organic-rich Marcellus Shale lithofacies

by support vector machine classifier in the Appalachian basin: Computers and

Geosciences, 64, 52-60.

West, P. B., R. S. May, E. J. Eastwood, and C. Rossen, 2002, Interactive seismic facies

classification using textural attributes and neural networks: The Leading Edge, 21, 1042-

1049.

Wong, K. W., Y. S. Ong, T. D. Gedeon, and C. C. Fung, 2005, Reservoir characterization using

support vector machines: Computational Intelligence for Modelling, Control and

Automation, 2005 and International Conference on Intelligent Agents, Web Technologies

and Internet Commerce, International Conference on, 2, 354-359.

Yu, S., K. Zhu, and F. Diao, 2008, A dynamic all parameters adaptive BP neural networks model

and its application on oil reservoir prediction: Applied Mathematics and Computation, 195,

66-75.

Zhang, B., T. Zhao, X. Jin, and K. J. Marfurt, 2015, Brittleness evaluation of resource plays by

integrating petrophysical and seismic data analysis (accepted by Interpretation).

Zhao, T., V. Jayaram, K. J. Marfurt, and H. Zhou, 2014, Lithofacies classification in Barnett Shale

using proximal support vector machines: 84th Annual International Meeting, SEG,

Expanded Abstracts, 1491-1495.

48

Zhao, T. and K. Ramachandran, 2013, Performance evaluation of complex neural networks in

reservoir characterization: applied to Boonsville 3-D seismic data: 83rd Annual

International Meeting, SEG, Expanded Abstracts, 2621-2624.

Zhao, B., H. Zhou, and F. Hilterman, 2005, Fizz and gas separation with SVM classification: 75th

Annual International Meeting, SEG, Expanded Abstracts, 297-300.

49

LIST OF FIGURE CAPTIONS

Figure 1. Classification as applied to the interpretation of seismic facies (Modified from Duda et

al., 2000).

Figure 2. Cartoon illustration of a k-means classification of 3 clusters. (a) Select 3 random or

equally spaced, but distinct seed points, which serve as the initial estimate of the vector means of

each cluster. Next, compute the Mahalanobis distance between each data vector and each cluster

mean. Then color code or otherwise label each data vector to belong to the cluster that has the

smallest Mahalanobis distance. (b) Recompute the means of each cluster from the previously

defined data vectors. (c) Recalculate the Mahalanobis distance from each vector to the new cluster

means. Assign each vector to the cluster that has the smallest distance. (d) The process continues

until the changes in means converge to their final locations. If we now add a new (yellow) point,

we will use a Bayesian classifier to determine into which cluster it falls (Figure courtesy of Scott

Pickford).

Figure 3. (a) A distribution of data points in 3-dimensional attribute space. The statistics of this

distribution can be defined by the covariance matrix. (b) k-means will cluster data into a user-

defined number of distributions (4 in this example) based on Mahalanobis distance measure. (c)

The plane that best fits these data is defined by the first two eigenvectors of the covariance matrix.

The projection of the 3D data onto this plane provides the first two principle components of the

data as well as the initial model for both our SOM and GTM algorithms. (d) SOM and GTM

deform the initial 2D plane into a 2D “manifold” that better fits the data. Each point on the

deformed 2D manifold is in turn mapped to a 2D rectangular “latent” space. Clusters are color-

coded or interactively defined on this latent space.

50

Figure 4. (a) K grid points uk defined on a L-dimensional latent space grid are mapped to K grid

points mk lying on a non-Euclidean manifold in N-dimensional data space. In this paper, L=2 and

will be mapped against a 2-dimensional color bar. The Gaussian mapping functions are initialized

to be equally spaced on the plane defined by the first two eigenvectors. (b) Schematic showing

the training of the latent space grid points to a data vector aj lying near the GTM manifold using

an expectation maximization algorithm. The posterior probability of each data vector is calculated

for all Gaussian centroids points mk and are assigned to the respective latent space grid points uk.

Grid points with high probabilities are displayed as bright colors. All variables are discussed in

Appendix.

Figure 5. Cartoon of a linear SVM classifier separating black from white data vectors. The two

dashed lines are the margins defined by support vector data points. The red decision boundary falls

midway between the margins, separating the two clusters. If the data clusters overlap, no margins

can be drawn. In this situation the data vectors will be mapped to a higher dimensional space where

they can be separated.

Figure 6. Cartoon describing semi-supervised learning. Blue squares and red triangles indicate

two different interpreter defined classes. Black dots indicate unclassified points. In semi-

supervised learning, unclassified data vectors 1 and 2 are classified to be class “A” while data

vector 3 is classified to be class “B” during the training process.

51

Figure 7. (a) Cartoon showing a two-class PSVM in 2D space. Classes “A” and “B” are

approximated by two parallel lines that have been pushed as far apart as possible forming the

cluster “margins”. The red decision-boundary lies midway between the two margins. Maximizing

the margin is equivalent to minimizing (𝝎𝑇𝝎 + 𝛾2)1 2⁄ . (b) A two-class PSVM in 3D space. In

this case the decision-boundary and margins are 2D planes.

Figure 8. Cartoon show how one SVM can map two linearly inseparable problem into a higher

dimensional space in which they can be separated. (a) Circular classes “A” and “B” in a 2D space

cannot be separated by a linear decision-boundary (line). (b) Mapping the same data into a higher

3-dimensional “feature” space using the given projection. This transformation allows the two

classes to be separated by the green plane.

Figure 9. A map showing the location of the 3D seismic survey acquired over the Canterbury

Basin, offshore New Zealand. The black rectangle denotes the limits of the Waka-3D survey, while

the smaller red rectangle denotes the part of the survey shown in subsequent figures. Colors

represent the relative depth of the current seafloor, warm being shallower and cold being deeper.

Current seafloor canyons are delineated in this map, which are good analogs for the paleocanyons

in Cretaceous and Tertiary ages (Modified from Mitchell and Neil, 2012).

Figure 10. Time slice at t=1.88 s through the seismic amplitude volume. White arrows indicate

potential channel/ canyon features. The yellow arrow indicates a high amplitude feature. Red

arrows indicate relatively low energy, gently dipping area. AA’ denotes a cross section shown in

Figure 14.

52

Figure 11. Time slice at t=1.88 s through peak spectral frequency co-rendered with peak spectral

magnitude that emphasizes the relative thickness and reflectivity of the turbidite system and

surrounding slope fan sediments into which it was incised. The two attributes are computed using

a continuous wavelet transform algorithm. The edges of the channels are delineated by Sobel filter

similarity.

Figure 12. Time slice at t=1.88 s through the GLCM homogeneity attribute co-rendered with

Sobel filter similarity. Bright colors highlights areas with potential fan sand deposits.

Figure 13. Time slice at t=1.88 s through the co-rendered shape index, curvedness, and Sobel

filter similarity. The shape index highlights incisement, channel flanks, and levees providing an

excellent image for interactive interpreter-driven classification. However, the shape index

dominates the unsupervised classifications, highlighting valley and ridge features and minimizing

more planar features of interest in the survey.

Figure 14. Vertical slices along line AA’ (location shown in Figure 10) through a) seismic

amplitude, b) seismic amplitude co-rendered with peak spectral magnitude and peak spectral

frequency, c) seismic amplitude co-rendered with GLCM homogeneity, and d) seismic amplitude

co-rendered with shape index and curvedness. White arrows indicate incised channel and canyon

features. The yellow arrow indicates at a high amplitude reflector. Red arrows indicate relatively

low amplitude, gently dipping areas.

Figure 15. Time slice at t=1.88 s through k-means classification volume with K=16. White arrows

indicate channel-like features. Yellow arrows indicate high amplitude overbank deposits. Red

arrows indicate possible slope fans. The edges of the channels are delineated by Sobel filter

similarity.

53

Figure 16. Time slice at t=1.88 s through k-means classification volume with K=256. The

classification result follows the same pattern as K=16 but is more chaotic since the classes are

computed independently and are not constrained to fall on a lower dimensional manifold. Note the

similarity between clusters of high amplitude overbank (yellow arrows) and slope fan deposits (red

arrows) which were separable in Figure 15.

Figure 17. Time slice at t=1.88 s of the first two principle components plotted against a 2D color

bar. These two principal components serve as the initial model for both the SOM and GTM images

that follow. With each iteration, the SOM and GTM manifolds will deform to better fit the natural

clusters in the input data.

Figure 18. Time slice at t=1.88 s through an SOM classification volume using 256 clusters. White

arrows indicate channel-like features. Combined with vertical sections through seismic amplitude,

we interpret overbank deposits (yellow arrows), crevasse splays (orange arrows), and slope fan

deposits (red arrows). The data are mapped to a 2D manifold initialized by first two principle

components and are somewhat more organized than the k-means image shown in the previous

figures.

Figure 19. Time slice at t=1.88 s through crossplotting GTM projection 1 and 2 using a 2D

colorbar. White arrows indicate channel-like features, yellow arrows overbank deposits, and red

arrows slope fan deposits. The blue arrow indicates a braided channel system that can be seen on

PCA but cannot be identified from k-means or SOM classification maps. The color indicates the

location of the mean probability of each data vector mapped into the 2D latent space.

54

Figure 20. The same time slice through the GTM projections shown in the previous image but

now displayed as four seismic facies. To do so, we first create two GTM “components” aligned

with the original first two principal components. We then pick four colored polygons representing

four seismic facies on the histogram generated using a commercial crossplot tool, This histogram

is a map of the GTM posterior probability distribution in the latent space. The yellow polygon

represents overbank deposits, the blue polygon channels /canyons, the green polygon slope fan

deposits, and the red polygon “everything else”.

Figure 21. Time slice at t=1.88 s through co-rendered peak spectral frequency, peak spectral

magnitude, and Sobel filter similarity volumes. Seed points (training data) are shown with colors

for the picked four facies, blue indicating multistoried channels, yellow point bars and crevasse

splays, red channel flanks, and green slope fans. Attribute vectors at these seed points are used as

training data in supervised classification.

Figure 22. PNN errors through the training epochs. The neural network reaches its best

performance at epoch 42.

Figure 23. Confusion tables for the same PNN shown in Figure 21. From these tables we find the

training correctness to be 90%, the testing and cross-validation correctness to be 86% and 91%,

warranting a reliable prediction.

Figure 24. Time slice at t=1.88 s through the ANN classification result. White arrows indicate

channels/canyons. Yellow arrows indicate point bars and crevasse splays.

55

Figure 25. Time slice at t=1.88 s through SVM classification result. White arrows indicate more

correctly classified slope fans. Yellow arrow indicates crevasse splays. Red arrows show the

misclassifications due to possible acquisition footprint.

Figure 26. Time slice at t=1.88 s through inline dip component of reflector dip. Inline dip

magnitude provides a photo-like image of the paleocanyons.

Figure A1. Self-organizing maps (SOM) workflow.

Figure A2. Generative topographic mapping (GTM) workflow.

Figure A3. Artificial neural network (ANN) workflow.

Figure A4. Proximal support vector machine (PSVM) workflow.

56

LIST OF TABLE CAPTIONS

Table 1. Attribute expressions of seismic facies.

Table 2. Classification settings and runtimes.

Table A1. List of shared mathematical symbols.

Fig. 1

Sensing

Segmentation

Feature extraction

Classification

Post-processing

Decision

Conventional interpretation of geologic

formations; geometric attribute delineation of

channel-levee complexes, fault blocks, ...

Reduce seismic wiggle samples to a few key

measures (AI, spectral components, curvature, …)

Interpreter-driven classification

using geologic models

Interpreter evaluation of alternative

working hypotheses

Validation. Do we need to

modify selected attributes?

Seismic data acquisition; well logging

Fig. 2

a) b)

c) d)

a)

c) d)

b)

Fig. 3

Fig. 4 a)

b)

N

𝑎1

𝑎1

𝑎2

𝑎2

𝑎3

𝑎3

Fig. 5

0

Support vectors

Class AClass B

𝑎1

𝑎2

Fig. 6

Class AClass BUnknown

1

2

3

0

𝑎2

𝑎1

Fig. 7

(a) (b)

Fig. 8

(a) (b)

𝑥2 + 𝑦2 = 1

𝑥2 + 𝑦2 = 2

A:

B:

(𝑥, 𝑦)

(𝑥, 𝑦, 𝑥2 + 𝑦2)

Denotes “A”Denotes “B”

Fig. 9

17

0°

30

’ E

17

3°

00

’ E

45°30’ S

46°30’ S

N

0 30 km

Fig. 100 5 km

Positive

Negative

Amplitude

A

A’

t = 1.88 s

N

Histo

gram

Fig. 110 5 km

High

Low

Peak Frequency

Peak Spectral

Magnitude

t = 1.88 s

N

Low

High

0 100%

Op

acity

0

Similarity

0.3

1

100%

Op

acity

Histo

gram

Histo

gram

Fig. 120 5 km

0.7

0.1

GLCM

homogeneity

t = 1.88 s

NSimilarity

0.3

1

100%

Op

acity

0%H

istogram

Fig. 130 5 km

t = 1.88 s

N

0

Similarity

0.3

1

100%

Op

acity

dome

bowl

Shape index

Curvedness

0

0.04

0 100%

Op

acity

Histo

gram

Histo

gram

ridge

valley

plain

Fig. 14

1.7

2.1

Tim

e (s

)

A A’NE

Am

p

+

-

2 km

Z scale 1:20

1.7

2.1

Tim

e (s

)

A A’NE

1.7

2.1

Tim

e (s

)

A A’NE

Am

p

+

-

2 km

Z scale 1:20

1.7

2.1

Tim

e (s

)

A A’NE

Op

acit

y

100%0.1

0.7

GLCM homogeneity

Op

acit

y

100%Low

High

Peak frequency

Op

acit

y

100%Low

High

Peak magnitude

Op

acit

y

100%Bowl

Dome

Shape index

Op

acit

y

100%0

0.04

Curvedness

(a) (b)

(c) (d)

Op

acit

y

100%

Am

p

+

-

2 km

Z scale 1:20

Op

acit

y

100%

Am

p

+

-

2 km

Z scale 1:20

Op

acit

y

100%

Fig. 150 5 km

16

1

k-means

clusters

t = 1.88 s

N

0

Similarity

0.3

1

100%

Op

acity

Fig. 160 5 km

256

1

k-means

clusters

t = 1.88 s

N

0

Similarity

0.3

1

100%

Op

acity

Fig. 170 5 km

t = 1.88 s

N

2D colorbar

Pri

nci

ple

com

ponen

t 1

Similarity

0.3

1

100%

Op

acity

Principle component 2

Pri

nci

ple

com

ponen

t 1

2D histogram

Principle component 2

Fig. 180 5 km

256

1

SOM

clusters

t = 1.88 s

N

0

Similarity

0.3

1

100%

Op

acity

Fig. 190 5 km

2D colorbar

1

10

0

GT

M l

aten

t ax

is 2

GTM latent axis 1

GT

M l

aten

t ax

is 2

2D histogram

GTM latent axis 1

t = 1.88 s

N

Similarity

0.3

1

100%

Op

acity

Fig. 200 5 km

GTM latent axis 1G

TM

lat

ent

axis

2

2D histogram

GTM latent axis 1

GT

M l

aten

t ax

is 2

2

1

4

3

Cluster selection

t = 1.88 s

N

Similarity

0.3

1

100%

Op

acity

Fig. 21

Facies index

Blue Channel thalweg

Yellow Overbank deposit

Red Channel flanks

Green Fan deposit

0 5 km

High

Low

Peak Frequency

Peak Spectral Magnitude

t = 1.88 s

N

Low

High

0 100%

Op

acity

Similarity

0.3

1

0 100%

Op

acity

Facies index

Blue Channel stories

YellowPoint bars and

crevasse splays

Red Channel flanks

Green Slope fans

Fig. 22

Fig. 23

0 5 km

t = 1.88 s

NFig. 24

Facies index

Blue Channel stories

YellowPoint bars and

crevasse splays

Red Channel flanks

Green Slope fans

0

Similarity

0.3

1

100%

Op

acity

0 5 km

t = 1.88 s

NFig. 25

Facies index

Blue Channel stories

YellowPoint bars and

crevasse splays

Red Channel flanks

Green Slope fans

0

Similarity

0.3

1

100%

Op

acity

Fig. 260 5 km

+10°Inline dip

-10°

t = 1.88 s

N

His

togra

m

r < σ

Yes

Shrink

neighborhood σ

and run training

loop

Deform neighboring

prototype vectors mk

Project final mk against onto

the 2D latent space grid uk

Sammon projection

Color the projected points uk

with 1D or 2D colorbar and

color-code corresponding

data points

Final Seismic facies

volumeMore

iterations

?

Yes

Final manifold

Prototype Vectors mk

Find nearest prototype vector

mk to each data point aj.

Assign aj to class k.

No

No

Input multi-attribute

data aj

Normalized data

vector aj

Randomly select the

data vector aj find the

Best match with the

prototype vectors mk

Initialize the 2D

latent space grid uk

Z-score

Initialize manifold

prototype vectors mk to

span the first two

principal components in

the data-space

Update mk

Training Loop

Fig. A1

Compute Gaussian with

centers at each mk

Initialize W

and β

Map the latent space vectors

uk to data space manifold

vectors mk=WΦ(uk)

Calculate the posterior

probabilities Rkj

Initialize manifold mk to span

the first two principal

components in the data-space

Δβ

small?

No Yes

Initialize the 2D latent space

grid uk and interpolation

function Φ(uk)

Calculate the new posterior

responsibility, Rkj

Update W and β

Updated W

and βProject the final mean

posterior responsibilities <Rkj>

onto the the 2D Latent space

using crossplot software

Final manifold mk

Draw polygons around

desired clusters and color-

code corresponding data

points

Input multi-attribute

data aj

Final Seismic facies

volume

Z-score

Fig. A2

Input multi-attribute

data aj

Output classes

Classified interpreter seed

picks

Attribute extraction

Training data

(classed attribute vectors)

jyje

r

1

1

N

n

njnj awy0

Form perceptrons

Update

W

Training

error

small?

YesNo

Final weights, W

jyje

r

1

1

N

n

njnj awy0

Apply perceptrons

Validation data

(classed attribute vectors)

Validation

OK?

Accept

Classification

Yes

Reject

Classification

r

y0.0

0.5

1.0

0.0

+1.0-1.0

yes

no

No

Reselect

input

attributes

and seed

picks

Fig. A3

A B C D

A 0.3 -1.2 2.3

B -0.3 0.8 -1.1

C 1.2 -0.8 -1.9

D -2.3 1.1 1.9

Example of a classification factor table

Evaluate the binary PSVM classification factor, f(q;p), of

the current pilot class, p, against all remaining active

classes, q

All f(q;p)

> 0? Yes

Choose initial pilot

class, p

Initialize all classes

to be active

No

For training samples of Q classes, generate

Q(Q-1)/2 binary classifiers for every two classes

Set old pilot, p, to be inactive

Set new p to be that q that gives min[f(q:p)]

Input multi-attribute

data aj

Assign the current pilot

class, p, to current voxel

Fig. A4

pq

Table 1

Facies Appearance to Interpreter Attribute Expression

Levee

Structurally highStronger dome or ridge shape structural components

locally continuousHigher GLCM homogeneity; lower GLCM entropy

Higher amplitude Dome or ridge shape component

Possibly thicker Lower peak spectral frequency

Channel thalwegsShale-filled with negative compaction

Stronger bowl or valley shape structural components; higher peak spectral frequency

Sand-filled with positive compactionStronger dome or ridge shape structural components; lower peak spectral frequency

Channel flanks Onlap onto incisement, canyon edges Higher reflector convergence magnitude

Gas-charged sands High amplitude, continuous reflectionsHigher GLCM homogeneity; lower GLCM entropy; high high peak magnitude

Incised floodplain Erosional truncationHigher reflector convergence magnitude, Higher curvedness

Floodplain

Lower amplitude Lower spectral magnitude

Higher frequency Higher peak spectral frequency

Continuous Higher GLCM homogeneity; lower GLCM entropy

Near planar eventsLower amplitude structural shape components; lower reflector convergence magnitude

Slumps chaotic reflectivityHigher reflector convergence magnitude; higher spectral frequency; lower GLCM homogeneity; higher GLCM entropy

Table 2

Algorithm Number of classes MPI processors*Dataset size (samples) Runtime (sec.)

Training Total TrainingApplying to the entire dataset

Total

k-means 16 50 809,600 101,200,000 65 20 85

k-means 256 50 809,600 101,200,000 1060 70 1130

SOM 256 1 809,600 101,200,000 4125 6035 10160

GTM - 50 809,600 101,200,000 9582 1025 10607

ANN+ 4 1 437 101,200,000 2 304 306

SVM 4 50 437 101,200,000 24 12092 12116

*SOM is not run under MPI in our implementation. ANN is run using MATLAB® and is not under MPI. All other three are run under MPI when applying the model to the entire dataset.+ANN is implemented using MATLAB® toolbox.

Variable Name Definitiion

n, N attribute index and number of attributes

j, J voxel (attribute vector) index and number of voxels

k, K manifold index and number of grid points

aj the jth attribute data vector

p matrix of principle components

C attribute covariance matrix

µn mean of the nth attribute

λm, vm the mth eigenvalue and eigenvector pair

mk

the kth grid point lying on the manifold (prototype vector for SOM, or Gaussian center for GTM)

uk the kth grid point lying on the latent space

rjk

the Mahalanobis distance between the jth data vector and the kth cluster center or manifold grid point

I Identity matrix of dimension defined in the text

Table A1